Shopify's Mikhail Parakhin discusses internal AI tooling, SimGym customer simulation, and deployment bottlenecks

Mikhail Parakhin, who previously led a Microsoft business unit spanning Windows, Edge, Bing, and advertising, joined Shopify as CTO and has become more vocal about the company's internal AI infrastructure. In a recent Latent Space podcast episode, Parakhin discussed how a 20-year-old, $200 billion software company navigates the shift toward AI-first operations.

The conversation centered on three major Shopify systems: Tangle, a framework for reproducible ML and data workflows with built-in content-addressed caching; Tangent, an automated research loop for optimizing search, themes, and prompt compression; and SimGym, a customer behavior simulator that uses historical transaction data to model merchant and buyer trajectories. According to Parakhin, these systems become more powerful when integrated, enabling non-ML engineers and product managers to run experiments without building custom infrastructure.

Parakhin argued that the bottleneck in AI-assisted development has shifted from code generation to review, CI/CD, and deployment stability. While AI models produce code faster, he noted that this can paradoxically increase production bugs if review processes and test coverage do not scale correspondingly. He emphasized that Shopify built its own pull request review flow because existing off-the-shelf tools miss domain-specific validation requirements.

On metrics and token budgets, Parakhin suggested that raw token counts—a measure promoted by Nvidia CEO Jensen Huang—are directionally useful but frequently misapplied. He contended that critique loops, stronger models, and spending more on review than on generation are the real constraints in agent-driven development. Shopify is also experimenting with Liquid AI, a non-transformer architecture, for low-latency operations like query understanding and catalog lookups, though the company remains pragmatic about model selection.

The interview also touched on SimGym's competitive advantage: the system only works reliably when trained on real customer behavior histories. Parakhin described how it evolved from A/B testing frameworks to counterfactual analysis that can recommend specific changes to a single live storefront—suggesting discounts, campaigns, or notifications based on modeled outcomes. This approach requires infrastructure spanning multimodal models, browser farms, and distillation pipelines.